You are on page 1of 5

ISSN 2320 - 2602

Volume 8 No.3, March 2019

International Journal of Advances in Computer Science and Technology
Sherafgan Kattak, International Journal of Advances in Computer Science and Technology, 8(3), March 2019, 16 - 20
Available Online at

ABSTRACT process and it is an extremely difficult task to get, up to the

mark segments [8,9]. Various segmentation algorithms i.e.
High-Resolution Imagery is futile until its spectral, spatial, Quadtree Segmentation, Chess-board segmentation, and
and contextual properties are utilized. Object-based Multiresolution segmentation are available in eCognition
classification makes it possible with the help of powerful (Definiens cognition network technology). Among all, the
image segmentation (Multiresolution segmentation) Multiresolution segmentation is preferred for extraction of
algorithm. With the single segmentation level, it is difficult to hidden features i.e. vegetation in barren hill and
extract all available features in a digital image. This research encroachment of green land. In OBC, segmentation is
aimed to find the ideal parameter for extracting all features. directly proportional to better classification and feature
And used those features for classifying the images (with the extraction. On the same segmented HRI, two supervised
difference of a decade period) using two supervised classifiers have been correlated; one is Fuzzy Nearest
classifiers, Fuzzy Nearest Neighbor and Support Vector neighbor (NN) and other is Support Vector Machine (SVM)
Machine.Classificationresults showed that Scale parameter in in terms of their classification accuracy and kappa statistics.
multiresolution segmentation is directly proportional to the This research is aiming to use eCognition built-in image
expandability of the image features. Moreover, it also segmentation algorithm (Multiresolution segmentation) for
showed that urban land has been increased in a decade extracting different features, classification of different land
period, while drastic reduction has occurred in green land. cover classes and the change occurred in a decade period of
Correlating between two classifiers, SVM performed better the same HRI.
with less execution time and higher overall accuracy of 0.98
and kappa statistics. 2. IMAGE SEGMENTATION
Of 0.97.
Image segmentation is a process of segregating the entire
Key words: Fuzzy Nearest Neighbor, Image Segmentation, digital image into segments and to classify these segments
Multiresolution Segmentation, Support Vector Machine. using classification methods [10]. Each segment is the
combination of identical pixels in respect of texture, color
and shape [11]. It is mostly used for identifying borders and
1.INTRODUCTION curves in an image. Several approaches are brought under
consideration based on its specific requirement, e.g. for
Currently, High-Resolution Imagery (HRI) is available to finding the edges; edge detection techniques and for feature
every researcher in remote sensing society. Having high extraction; region-based segmentation is used [12].
spatial and spectral resolution [1], it can efficiently be used
for generating precise land cover maps and extraction of 2.1 Multiresolution Segmentation
different features. In past literature [2-3] traditional
classification method reckons only on the pixel and its Multiresolution segmentation (MRS) is a built-in image
spectral values; which result is deplorable [4]. For maximum segmentation algorithm in eCognition. It starts working on
output, Object-Based Classification(OBC) is considered a considering a single pixel as an object along with its
worthy tool [5]. However, feature selection and integration of neighboring pixels and merges it based on the predefined
training samples are still a critical task [6]. In OBC, image criteria i.e. homogenous color, homogenous shape, texture
objects are formed using the segmentation process, which and intensity [13] In OBC, these merged pixels are known as
merges homogeneous pixels in terms of shape, color, texture, image objects. As image objects can be utilized for feature
contextual information and their interrelationship. These extraction and land cover classification, so MRS can be
image objects take part in further image analysis and categorized as region-basedsegmentation. The size ofimage
classification [7]. Moreover, segmentation is an iterative object is directly proportional to the user defined threshold,
Sherafgan Kattak, International Journal of Advances in Computer Science and Technology, 8(3), March 2019, 16 - 20

the Scale Parameter (SP); the higher SP leaves the algorithm

to repeat several times to make larger image objects. Other
parameters are shape(compactness, smoothness) andcolor.
[14]. Various levels of segmentation process are needed till
the findingof appropriate image objects; especially in feature


The focused area of this research is the urbanized area of
Peshawar, Pakistan.SPOT-5 satellite imagery has been
acquired from Pakistan Space Upper Atmosphere and
Research Commission (SUPARCO) for the year 2005 and
2015 for more than 40 square kilometers. These imageshave Figure 1: Multiresolution segmentation result for Level1 (SPOT-5
four bands (Red, Green, Blue, and NIR) with 10m spatial 2005)

4.1 Segmentation Measures
Each SPOT-5 HRI (2005 and 2015) is processed in a separate
eCognition project. And the same techniques, already
discussed in section II are applied. According to the ground
truth data, the total five featured classes including (urban
land, green land, barren hill, bare land, and wetland) are
identified. For these five classes, the images are segmented in
three levels also shown in the table.1 but the segmentation
result was similar for level 2 and level 3. This is because of
the maximum capacity of the algorithm to merge the pixels.
The segmentation results in the figures 1,2 and figures 3, 4
for 2005 and 2015 respectively, clearly show that the objects
in level2 have large sizes as compared to the objects in
Figure 2: Multiresolution segmentation result for Level2 (SPOT-5

Table 1: Multiresolution segmentation parameters for Level 1,2 and

3 respectively
Scale Shape
Levels Parameter Color
Algorithm Compact/Smooth
Level1 16 0.8 0.4/ 0.2
Algorithm Level2 40 0.7 0.7/0.3
Level3 80 0.6 0.8/0.4

Figure 3: Multiresolution segmentation result for Level1 (SPOT-5


Sherafgan Kattak, International Journal of Advances in Computer Science and Technology, 8(3), March 2019, 16 - 20

Figure 4: Multiresolution segmentation result for Level2 (SPOT-5

2015) Figure 6: Classification of (SPOT-5 2015) image by Nearest
4.2 Supervised Classifiers

In implementation, the second phase is to classify the

segmented images of the first phase into the five land cover
classes i.e. urban land, green land, barren hill, bare land and
wetland. For classification, two supervised classifiers fuzzy
NN and SVM are utilized. NNworks on mean of two
neighbors in the entire image [15]. While SVM works on
hyperplane and kernel [16].Classification results in figures 5,
6 and figures 7, 8 for 2005 and 2015 of NN and SVM are
shown respectively. Urban land is displayed in red, bare land
in yellow, barren hill in white, green land in green and
wetland in blue color.

Figure 7: Classification of (SPOT-5 2005) image by Support

Vector Machine

Figure 5: Classification of (SPOT-5 2005) image by Nearest


Figure 8: Classification of (SPOT-5 2015) image by Support

Vector Machine

Sherafgan Kattak, International Journal of Advances in Computer Science and Technology, 8(3), March 2019, 16 - 20


Confusion matrix also known as error matrix and Kappa

statisticsalso known as Cohen’s kappa are used to calculate
the accuracy. The confusion matrices are calculated in every
classifier for every image also shown in table 2.The detailed
misclassification is also generated on every execution of the
classifier, but those misclassifications are removed by
incorporating user ground truth experience into sample
selection process. Here overall accuracy and kappa
statisticsare used to correlate both classifiers. In both
classifiers, Wetland feature has extracted normally for any
segmentation level. The green land area is expanded, coversa
large area in 2005 image, so high SP value will result in large
objects. Thus, for extracting expanded objects, high SP
values are required which is also shown in figure 5 and 7. In Figure 9: Graphical representation of difference in objects with
2015, urban land is expanded, thus in level 2 of 2015, also respect to time from 2005 to 2015 (NN)
shown in figure 6 and 8, the urban land is classified in a
change profound manner covering large area. Comparing
with the ground truth data, the huge change occurred in a
decade period in Peshawar, also shown in table 3 and figures
9 and 10. Correlating both classifiers, the SVM values are

Table: 2: Error Matrix (Overall Accuracy and Kappa Statistics of

NN and SVM)
Years 2005 2015


Overall Accuracy 0.962 0.978 0.96 0.983

Kappa Statistic 0.961 0.972 0.96 0.98

Figure.10: Graphical representation of difference in objects with
respect to time from 2005 to 2015(SVM)

Table: 3: Difference in objects with respect to time from 2005 to
2015 (NN and SVM) In this research, the object-based classification on the High-
resolution imagery has been utilized. The Image
Years segmentation and particularly Multiresolution segmentation
N S N S N S N S N S has implemented on SPOT-5 Satellite imagery. It is pertinent
to note that the image segmentation was worth using for HRI
2005 66 68 20 21 198 182 144 170 425 413
due to the utilization of all its properties in a good manner.
2015 53 55 18 18 329 289 112 139 287 298 Moreover, the image features have a direct connection with
the scale parameter of the MRS. And multiple features can be
- - - -
+39 + 35 -28 -22 -48 -38 extracted by varying the scale parameter value along with the
19 23 11 16
ence shape and compactness. Between the two classifiers, SVM
performed better, in its execution time and results. The urban
land has drastically increased while the green land has


1. Jiang, N., et al. "Semi-automatic building extraction from

high resolution imagery based on segmentation." 2008
International Workshop on Earth Observation and Remote
Sensing Applications. IEEE, 2008.
Sherafgan Kattak, International Journal of Advances in Computer Science and Technology, 8(3), March 2019, 16 - 20

2. Ahmed et al., "Comparing Pixel-Based Classifiers for 14. Darwish, K. Leukert, and W. Reinhardt, "Image
Detecting Tobacco Crops in North-West Pakistan," in Recent Segmentation for The Purpose of Object-Based Classification
Advances in Space Technologies (RAST), 2015 7th " in Geoscience and Remote Sensing Symposium, 2003.
International Conference on, 2015, pp. 211-216: IEEE. IGARSS'03. Proceedings. 2003 IEEE International, 2003, vol. 3, pp. 2039-2041: Ieee.
15. Iman, Paryudi,Alternative Design Exploration using K
3. Expansion and Land Use Changes In Akure, Nigeria, Using
Nearest Neighbor Technique andSemantic Web Technology
Remote Sensing And Geographic Information System (GIS)
in an Energy Simulation Tool, International Journal of
Techniques," vol. 4, no. 9, pp. 533-541, 2011.
Advances in Computer Science and TechnologyVolume 2,
4. Xiaoxia, Sun, Zhang Jixian, and Liu Zhengjun. An object- No.10, October 2013.
oriented classification method on high resolution satellite
16. Tzotsos and D. Argialas, "Support Vector Machine
Classification for Object-Based Image Analysis," in Object-
Based Image Analysis: Springer, 2008, pp. 663-677.
5. T. Blaschke, "Object Based Image Analysis for Remote
Sensing," ISPRS Journal of Photogrammetry and Remote
Sensing, vol. 65, no. 1, pp. 2-16, 2010.

6. S. Laliberte, J. Koppa, E. L. Fredrickson, and A. Rango,

"Comparison of nearest neighbor and rule-based decision tree
classification in an object-oriented environment," in
Proceedings of IEEE international Geoscience and Remote
Sensing Symposium, Denver, Colorado, USA, 2006, pp.

7. Wong, T. H., et al. "Feature extraction based on object-

oriented analysis." Proceedings of ATC 2003 Conference.
Vol. 2021. 2003.
8. Dey, Vivek, Yun Zhang, and Ming Zhong. "A review on
image segmentation techniques with remote sensing
perspective." (2010): 31-42.
9. Wen Zhou, Dongping Ming, Lu Xu, Hanqing Bao, and Min
Wang, “Stratified Object-Oriented Image Classification
Based on Remote Sensing Image Scene Division,” Journal of
Spectroscopy, vol. 2018, Article ID 3918954, 11 pages,
10. Ajay Kumar Thakur, Madhusudan, Image Segmentation: A
Comparative study, International Journal of Advances in
Computer Science and Technology, Volume 5, No.6, June
2016, 64-71
11. Kaur, Harmanpreet. "Review of remote sensing image
segmentation techniques." International Journal of Advanced
Research in Computer Engineering & Technology
(IJARCET) 4.4 (2015): 1667-1674.
12. U. C. Benz, P. Hofmann, G. Willhauck, I. Lingenfelder, and
M. Heynen, "Multi-Resolution, Object-Oriented Fuzzy
Analysis of Remote Sensing Data For GIS-Ready
Information," ISPRS Journal of Photogrammetry and Remote
Sensing, vol. 58, no. 3-4, pp. 239-258, 2004.

13. Trimble, "eCognition-ReferenceBook," 2012.